Pydantic AI re-sends the whole run history every step, so token spend goes quadratic by design

A four-step run sends ten copies of the history in total; a twelve-step run sends seventy-eight ^[1]. That is 7.8 times the payload, against the four times the dev.to post uses as its worked illustration ^[10][2]. The exponent is structural. The coefficient is not, and it lives in two places the run graph does not control.

The first is caching. The post's own pricing buckets put cache-write at roughly 1.25 times input and cache-read at about a tenth ^[11]. An append-only prefix that never changes is the friendliest possible input to a prompt cache, so where the provider actually serves it, the growth continues at something like a tenth the rate ^[3]. Prefix hit rate, not step count, is then the number worth wiring to a dashboard.

The second is what gets copied. The post quotes `messages[:] = _clean_message_history(ctx.state.message_history)` and does not say what that call strips ^[5]. Read strictly, the evidence supports history accumulating and being sliced whole into each step ^[2][4]; it does not establish that nothing ever leaves the wire before the request goes out.

The escape hatch is real, and it sits outside the loop. Every run method accepts a `message_history` argument, documented as the conversation so far, and `result.all_messages()` hands the transcript back ^[6], so between turns you choose what to replay. Within a single deep tool loop the author's own position is that the re-send is inherent to how tool calling works and true of every framework ^[7]. There is nothing to patch, then. What remains is a sizing question about tool outputs and loop depth, and that question belongs to whoever wrote the tools.

One source, and an interested one: the post closes by recommending the author's own npm package for turning real usage into a per-run dollar figure ^[11]. The mechanism does not lean on that, because it rests on quoted lines from `_agent_graph.py` in `pydantic_ai_slim` on main ^[14], which anyone can open and check against the version they actually installed. The 4x does lean on it, being an output of the tooling rather than of the arithmetic ^[10]. Take the exponent as documented and the multiplier as yours to measure.